Subscribe to our daily and weekly emails for the latest updates on AI. Exclusive content is also available. Learn More
AI is moving fast, and if your company doesn’t keep up, it will fall behind.
Two recent innovations are reshaping both the landscape and the developers’ perspective: DeepSeek R1 model release, and OpenAI’s Deep Research product. They’re working together to redefine the cost and accessibility for powerful reasoning models. This has been well-reported. It’s less well-known, however, how they will push companies to use techniques such as distillation, reinforcement learning (RL), supervised fine tuning (SFT) and retrieval augmented generation (RAG) in order to build smarter and more specialized AI apps.
Once the initial excitement over DeepSeek’s amazing achievements has subsided, developers and enterprise decision makers need to consider the implications for themselves. Here’s what these breakthroughs in AI mean for today’s AI developers, from pricing and performance to hallucination risk and the importance clean data.
Through distillation, we can achieve cheaper, transparent, and industry-leading models of reasoning.
With DeepSeek-R1, the headline is simple: it delivers an industry leading reasoning model for a fraction of OpenAI’s O1. DeepSeek is about 30 times cheaper than many closed models to run. And unlike many closed models that are opaque, DeepSeek provides full transparency of its reasoning steps. This means that developers can now build highly customizable AI models without breaking their bank — whether it’s through distillation, fine tuning or simple RAG implementaions.
In particular, distillation is a powerful tool. DeepSeek R1 can be used as a “teacher” model to create smaller models for specific tasks that inherit R1’s superior reasoning abilities. These smaller models are in fact the future of most enterprise companies. The full R1 reasoning can be too much. Companies may not take the necessary actions for their domain applications because they are thinking or too much.
“One thing that is not being discussed, especially in the mainstream media is that reasoning models don’t work that well for agents,” said Sam Witteveen. He is a machine learning developer who works with AI agents that orchestrate enterprise applications.
DeepSeek’s release included a distillation of its reasoning capabilities into a number smaller models. These models included open-source models such as Meta’s Llama and Alibaba’s Qwen families. In its paperdescribed. These smaller models can be optimized to perform specific tasks. This trend towards smaller, faster models to meet custom-built requirements will accelerate. Eventually, there will be armies.
We are moving into a world where people use multiple models. Witteveen said that people are not using the same model all the time. This includes the smaller, cheaper closed-source models from Google and OpenAI. “This means that models such as Gemini Flash, GPT-4o Mini and these really inexpensive models work really well for 80% use cases.”
If you are in a niche domain and have the resources, use SFT…
Enterprise companies have several options after the distillation step to ensure the model is ready for the application they need. If you work in a domain where there are no details on the web, or in books, which is what large language models (LLMs), typically, train on, you can inject your own domain-specific datasets with SFT. Ship container-building is an example, where specifications, protocol and regulations are not readily available.
DeepSeek demonstrated that you can achieve this with “thousands of” question-answer datasets. IBM engineer Chris Hay showed how others could put this into action by demonstrating how he fine-tuned his own small model using math-specific datasets in order to achieve lightning-fast answers — outperforming OpenAI’s o1 on similar tasks (View the hand-on video). Here
…and a bit of RL
Companies that want to train a model to be more aligned to specific preferences – for example, to make a customer service chatbot sound empathetic and concise – will also want to do a little RL. This is good for companies who want their chatbots to adapt their tone and recommendations based on feedback from users. Wharton AI Professor: As models become more and more capable, “personality”or the way they behave, will be increasingly important. Ethan Mollick on X said
The implementation of these SFT and RL measures can be difficult for companies. If you feed the model data from a specific domain area or tune it in a certain manner, it will suddenly become useless for tasks outside that domain.
RAG is a good option for most companies
RAG, however, is the safest and easiest way to go forward. RAG is an easy-to-use process that allows organizations the ability to use their own proprietary databases as a basis for their models. This ensures outputs are accurate and domain specific. An LLM searches vector and graph databases for information relevant to a user’s query. RAG processes are very good at identifying only the most relevant information.
Using this approach can also help counteract some of DeepSeek’s hallucination problems. According to a study, DeepSeek hallucinates 14 percent of the time, compared to only 8 percent for OpenAI’s o3 model. Vectara, a vendor who helps companies with RAG processes, has done a study.
The magic for most companies will be in the distillation of models and RAG. It is now so easy to do that even those with little or no data science or coding knowledge can do it. I downloaded the DeepSeek 1.5b Qwen model – the smallest – to fit it on my Macbook Air. I then loaded some PDFs of resumes from job applicants into a vector data base, and asked the model to scan the resumes to determine which applicants were qualified to work for VentureBeat. This took me 74 lines, which I essentially borrowed from Others are doing the same).
The Deepseek distilled models showed the thought process behind why it recommended or did not recommend each applicant. This was a level of transparency I would have never experienced before Deepseek released.
I demonstrated in my recent video on DeepSeek RAG how easy it is to implement RAG into practical applications for non-experts. Witteveen contributed to the discussion as well by explaining how RAG pipelines function and why enterprises increasingly rely on them rather than fine-tuning their models. ( Watch it here ().
OpenAI Deep Research: Extending RAG’s capabilities, but with caveats.
Whereas DeepSeek makes reasoning models more transparent and cheaper, OpenAI’s Deep Research represents an entirely different but complementary shift. It can take RAG up a notch by crawling the internet to create highly customized research. The output of the research can be used as input in RAG documents that companies can use.
This capability, also known as agentic RAG allows AI systems autonomously to seek out the best context across the internet. It brings a new dimension in knowledge retrieval and edification.
Open AI’s Deep Research is similar in many ways to Google’s Deep Research, Perplexity, and You.com. However, OpenAI tries to differentiate itself by claiming that its superior chain of thought reasoning makes it more accurate. These tools work as follows: A researcher from a company asks the LLM to gather all the available information about a particular topic and compile it into a well researched and cited report. The LLM will then ask the researcher to answer 20 more sub-questions in order to confirm that they have understood what was requested. The research LLM will then perform 10 or 20 web search to find the most relevant information to answer those sub-questions. It will then extract and present the knowledge in a useful manner. This innovation is not without its challenges. Amr Awadallah, CEO of Vectara, warned against relying on models such as Deep Research. Awadallah questions whether it is indeed more accurate. “It is not clear that this claim is true,” he noted. “We’re reading articles and posts on various forums that say no, they are still getting lots of hallucinations, and Deep Research’s only about as good at other solutions available on the market.” Awadallah stated that the grounding knowledge of a model must come from human-approved, verified sources to avoid cascading mistakes.
The cost curve is crashing. Why does this matter?
DeepSeek’s aggressive price reduction is the most immediate impact. Few in the tech industry anticipated that costs would drop over time. However, few predicted how quickly they would do so. DeepSeek has shown that powerful open models can be both efficient and affordable, allowing for widespread experimentation.
Awadallah stressed this point, noting the inference costs, which are about 1/30th the OpenAI o1 or O3 inference costs per token. Awadallah said that the margins that OpenAI and Google Gemini had been able to capture would now have to be reduced by at least 90%, as they cannot remain competitive with such high prices. Not only that, but those costs will continue going down. Anthropic CEO Dario Amedei recently stated that the cost to develop models continues to fall at around a Rates are increasing by 4x each year. The rate that LLM providers will charge to use them is also expected to continue to decrease.
I fully expect the cost will go to zero,” Ashok Srivastava said, CDO of Intuit. The company has been implementing AI in its tax- and accounting software, such as TurboTax, Quickbooks, with great success. “…and latency will be zero. These are just going to become commodity capabilities that we can use.
The cost reduction is not just a win, but also a signal to developers and enterprise users that AI innovation isn’t confined to large labs with billion dollar budgets. The barriers to entry are lower, and this is inspiring smaller companies and developers to experiment in previously unthinkable ways. Srivastava said that the models are so easy to use, even non-AI experts can use them.
DeepSeek’s disruption – Challenging the “Big AI” stronghold on models development
DeepSeek is most important for destroying the myth that major AI labs are only able to innovate. Over the years, companies such as OpenAI and Google have positioned themselves to be the gatekeepers for advanced AI. They spread the belief that only PhDs with large resources could create competitive models.
DeepSeek is changing that narrative. It has enabled a new generation of developers and enterprises to experiment and innovate, without needing billions of dollars in funding. This democratization is especially important in the post-training stage — like RL or fine-tuning – where the most exciting development are taking place.
DeepSeek exposed an AI fallacy — that only large AI labs and companies were capable of innovation. This fallacy forced many other AI builders into the background. DeepSeek has stopped that. It has inspired everyone to think of new ways to innovate.
The Data Imperative: Why clean, curated, data is the next step for enterprise companies.
Although DeepSeek and Deep Research are powerful tools, they depend on one crucial factor: Data Quality. Data quality has been a major theme for many years and has increased in importance over the last nine years. It’s now even more crucial with generative AI and DeepSeek’s disruption. Hilary Packer (CTO of American Express) highlighted this in an interview to VentureBeat. The data was, in truth, the aha! moment for us. The data is the key. You can select the best models in the world, but it’s the data that matters. Validation and accuracy is the holy grail of generative AI right now.”
Here, enterprises must focus their effort. It’s tempting for enterprises to chase the latest AI models and techniques. However, the foundation of every successful AI application is clean data that has been well-structured. The quality of the data you use will determine how accurate and reliable your models are, whether you’re using SFT, RAG or RL.
While many companies strive to perfect their entire data eco-system, it is not possible. Instead, businesses should focus their efforts on cleaning and curating critical portions of data to enable AI applications that provide immediate value.
In relation to this, many questions remain about the exact data used by DeepSeek to train its models, and this raises questions regarding the inherent bias in the knowledge stored in the model weights. This is no different than questions about other open-source model series, such as Meta’s Llama. Most enterprise users have found a way to fine-tune the models or ground them with RAG in a way that mitigates any biases. This has been enough to generate a serious momentum in enterprise companies towards accepting open source and even leading the way with open source.
There’s no doubt that many companies will use DeepSeek models regardless of the fears around the fact that this company is based in China. It’s also true, however, that many companies in highly-regulated industries like finance or healthcare will be cautious to use any DeepSeek models in any application that interacts directly with customers.
Conclusion: The future enterprise AI is open, affordable, and data-driven.
DeepSeek, OpenAI’s Deep Research, and OpenAI’s Deep Research represent more than just tools in the AI arsenal. They are signals of a fundamental shift in which enterprises will roll out massive amounts of purpose-built AI models, at a very affordable price, with high competence, and based on the company’s data and approach.
The message for enterprises is clear: the tools to build powerful domain-specific AI apps are at your fingertips. If you don’t use these tools, you risk falling behind. But the real success will come when you learn to curate your data and use techniques like RAG, distillation, and innovate beyond pre-training.
AmEx’s Packer said: Companies that get their data correct will lead the next wave of AI innovations.
Want to impress your boss? VB Daily can help. We provide you with the inside scoop about what companies are doing to maximize ROI, from regulatory changes to practical deployments.
Read our privacy policy
Thank you for subscribing. Click here to view more VB Newsletters.
An error occured.
